mitai-jinkendo/backend/workflow_executor.py

"""
Workflow Executor (Phase 4)

Führt Workflows mit conditional branching und path consolidation aus.

Phase 2: Sequential execution
Phase 3: Conditional branching, Logic Nodes, Fallback strategies
Phase 4: Join Nodes, Path consolidation

Konzept-Basis: konzept_workflow_engine_konsolidated.md
Anforderungsanalyse: anforderungsanalyse_umsetzungsplan.md (Phase 2-4)
"""
from typing import Dict, Any, List, Optional, Set
from datetime import datetime
import uuid
import logging
import json
from jinja2 import Environment, ChainableUndefined, TemplateError

from workflow_models import (
    WorkflowGraph, WorkflowNode, NodeExecutionState, ExecutionResult,
    NodeStatus, NormalizedSignal, FallbackStrategy, SignalStatus,
    EndNodeOutputMode
)
from workflow_engine import parse_workflow_graph, get_execution_order
from question_augmenter import (
    augment_prompt_with_questions,
    parse_question_augmentations_from_jsonb
)
from result_container_parser import parse_result_container
from normalization_engine import normalize_all_signals, normalize_decision_signal, load_question_catalog
from logic_evaluator import evaluate_logic_expression, resolve_signal_reference
from join_evaluator import evaluate_join_node as evaluate_join_node_core
from db import get_db, get_cursor

logger = logging.getLogger(__name__)


async def execute_workflow(
    workflow_id: Optional[str] = None,
    graph_data: Optional[Dict] = None,  # Phase 5: Direkt von ai_prompts.graph_data
    profile_id: str = None,
    variables: Dict[str, Any] = None,
    openrouter_call_func = None,  # Callback für LLM-Calls: async (prompt, model) -> str
    enable_debug: bool = False,
    progress_callback = None  # NEW: Optional callback für Progress-Updates: async (event_type, data) -> None
) -> ExecutionResult:
    """
    Führt einen Workflow aus (mit conditional branching und path consolidation).

    Phase 2: Linear execution in topological order.
    Phase 3: Conditional branching basierend auf logic nodes.
    Phase 4: Join nodes und path consolidation.
    Phase 5: Unterstützt graph_data direkt (aus ai_prompts, nicht workflow_definitions)

    Args:
        workflow_id: UUID des Workflows (legacy, für workflow_definitions Tabelle)
        graph_data: Workflow-Graph als Dict (NEU, für ai_prompts.graph_data)
        profile_id: UUID des Profils
        variables: Platzhalter-Werte (z.B. {"name": "Lars", ...})
        openrouter_call_func: async (prompt, model) -> str
        enable_debug: Debug-Modus

    Returns:
        ExecutionResult mit allen node_states

    Beispiel (Phase 5):
        >>> result = await execute_workflow(
        ...     graph_data={"nodes": [...], "edges": [...]},
        ...     profile_id="test-profile",
        ...     variables={"name": "Lars"},
        ...     openrouter_call_func=my_llm_func
        ... )
    """
    execution_id = str(uuid.uuid4())
    started_at = datetime.utcnow().isoformat()

    logger.info(f"Starting workflow execution: {execution_id}")

    # NEW: Progress-Callback für Start
    if progress_callback:
        await progress_callback("execution_started", {
            "execution_id": execution_id,
            "started_at": started_at
        })

    try:
        # 1. Lade Workflow-Definition
        if graph_data:
            # Phase 5: Graph direkt aus ai_prompts.graph_data (already a dict)
            graph_dict = graph_data
            logger.debug(f"Using provided graph_data")
        elif workflow_id:
            # Phase 0-4: Graph aus workflow_definitions Tabelle (legacy)
            with get_db() as conn:
                cur = get_cursor(conn)
                cur.execute(
                    "SELECT graph FROM workflow_definitions WHERE id = %s AND active = true",
                    (workflow_id,)
                )
                row = cur.fetchone()
                if not row:
                    raise ValueError(f"Workflow not found: {workflow_id}")

                graph_dict = row['graph']  # PostgreSQL JSONB returns dict
            logger.debug(f"Loaded graph from workflow_definitions: {workflow_id}")
        else:
            raise ValueError("Entweder workflow_id oder graph_data muss übergeben werden")

        # 2. Parse Graph
        graph = parse_workflow_graph(graph_dict)
        logger.debug(f"Parsed graph: {len(graph.nodes)} nodes, {len(graph.edges)} edges")

        # 4. Lade Question Catalog
        with get_db() as conn:
            catalog = load_question_catalog(conn)
        logger.debug(f"Loaded catalog: {len(catalog)} question types")

        # 5. Execute Nodes mit conditional branching (Phase 3)
        node_states: List[NodeExecutionState] = []
        context = {
            "variables": variables,
            "profile_id": profile_id,
            "node_results": {},  # Phase 3: Store full NodeExecutionState
            "active_edges": {},  # Phase 3: Track edge activation
            "graph": graph       # Phase 4: Needed for End Node to access questions
        }

        # Phase 3: BFS traversal mit edge activation
        start_node = next((n for n in graph.nodes if n.type == "start"), None)
        if not start_node:
            raise ValueError("No start node found in workflow")

        visited: Set[str] = set()
        queue = [start_node.id]

        while queue:
            node_id = queue.pop(0)

            if node_id in visited:
                continue

            visited.add(node_id)
            node = next(n for n in graph.nodes if n.id == node_id)

            # Prüfe ob Node aktiv ist (mindestens eine incoming edge aktiv)
            if not _has_active_incoming_edge(node, graph, context):
                logger.info(f"Skipping node {node_id}: no active incoming edges")
                node_states.append(NodeExecutionState(
                    node_id=node_id,
                    status=NodeStatus.SKIPPED,
                    error="no_active_incoming_edges",
                    started_at=datetime.utcnow().isoformat(),
                    completed_at=datetime.utcnow().isoformat()
                ))
                continue

            logger.info(f"Executing node: {node_id} (type: {node.type})")

            node_state = await execute_node(
                node=node,
                context=context,
                catalog=catalog,
                graph=graph,  # Phase 3: Needed for logic nodes
                openrouter_call_func=openrouter_call_func,
                enable_debug=enable_debug
            )

            node_states.append(node_state)
            context["node_results"][node_id] = node_state

            # NEW: Progress-Callback aufrufen (für SSE Streaming)
            if progress_callback:
                # Create a meaningful label for the node
                node_label = node.prompt_slug if hasattr(node, 'prompt_slug') and node.prompt_slug else f"{node.type.value}-{node_id[:8]}"
                await progress_callback("node_complete", {
                    "node_id": node_id,
                    "node_type": node.type,
                    "node_label": node_label,
                    "status": node_state.status.value,
                    "total_nodes": len(graph.nodes),
                    "completed_nodes": len([ns for ns in node_states if ns.status in [NodeStatus.EXECUTED, NodeStatus.SKIPPED]]),
                    "error": node_state.error if node_state.status == NodeStatus.FAILED else None
                })

            # Füge Nachfolger zur Queue hinzu
            outgoing_edges = [e for e in graph.edges if e.from_node == node_id]
            for edge in outgoing_edges:
                # Prüfe ob Edge aktiv ist (default: True, Logic Nodes setzen auf False)
                if context["active_edges"].get(edge.id, True):
                    queue.append(edge.to_node)

        # 6. Aggregiere Ergebnisse
        aggregated = aggregate_results(node_states, graph)

        # 7. Speichere Execution State
        completed_at = datetime.utcnow().isoformat()
        save_execution_state(
            execution_id=execution_id,
            workflow_id=workflow_id,
            profile_id=profile_id,
            node_states=node_states,
            status="completed",
            started_at=started_at,
            completed_at=completed_at
        )

        logger.info(f"Workflow execution completed: {execution_id}")

        # NEW: Progress-Callback für erfolgreiche Fertigstellung
        if progress_callback:
            await progress_callback("execution_complete", {
                "execution_id": execution_id,
                "status": "completed",
                "aggregated_result": aggregated,
                "total_nodes": len(node_states),
                "completed_nodes": len([ns for ns in node_states if ns.status == NodeStatus.EXECUTED]),
                "skipped_nodes": len([ns for ns in node_states if ns.status == NodeStatus.SKIPPED]),
                "failed_nodes": len([ns for ns in node_states if ns.status == NodeStatus.FAILED]),
                "completed_at": completed_at
            })

        return ExecutionResult(
            execution_id=execution_id,
            workflow_id=workflow_id or "N/A",  # Placeholder when graph_data is used directly
            status="completed",
            node_states=node_states,
            aggregated_result=aggregated,
            started_at=started_at,
            completed_at=completed_at
        )

    except Exception as e:
        logger.error(f"Workflow execution failed: {e}", exc_info=True)

        # NEW: Progress-Callback für Fehler
        if progress_callback:
            await progress_callback("execution_failed", {
                "execution_id": execution_id,
                "status": "failed",
                "error": str(e),
                "completed_at": datetime.utcnow().isoformat()
            })

        # Speichere Failed State
        completed_at = datetime.utcnow().isoformat()
        save_execution_state(
            execution_id=execution_id,
            workflow_id=workflow_id,  # Can be None when graph_data is used
            profile_id=profile_id,
            node_states=node_states if 'node_states' in locals() else [],
            status="failed",
            started_at=started_at,
            completed_at=completed_at,
            error=str(e)
        )

        return ExecutionResult(
            execution_id=execution_id,
            workflow_id=workflow_id or "N/A",  # ExecutionResult requires string, use placeholder
            status="failed",
            node_states=node_states if 'node_states' in locals() else [],
            aggregated_result={},
            started_at=started_at,
            completed_at=completed_at,
            error=str(e)
        )


async def execute_node(
    node,
    context: Dict[str, Any],
    catalog: Dict[str, Dict],
    graph: WorkflowGraph,  # Phase 3: Needed for logic nodes
    openrouter_call_func,
    enable_debug: bool = False
) -> NodeExecutionState:
    """
    Führt einen einzelnen Knoten aus.

    Args:
        node: WorkflowNode (aus graph.nodes)
        context: Execution context (variables, profile_id, node_results, active_edges)
        catalog: Question catalog
        graph: WorkflowGraph (für Logic Nodes)
        openrouter_call_func: LLM callback: async (prompt, model) -> str
        enable_debug: Debug mode

    Returns:
        NodeExecutionState

    Node Types:
        - start/end: No-op
        - analysis: Load prompt → augment → LLM → parse → normalize
        - logic: Evaluate condition → activate/deactivate edges (Phase 3)
        - join: Consolidate paths → merge results (Phase 4)
    """
    started_at = datetime.utcnow().isoformat()

    try:
        # Start Node: No-Op
        if node.type == "start":
            logger.debug(f"Node {node.id}: No-op (start)")
            return NodeExecutionState(
                node_id=node.id,
                status=NodeStatus.EXECUTED,
                started_at=started_at,
                completed_at=datetime.utcnow().isoformat()
            )

        # End Node: Output Generation (Phase 4)
        if node.type == "end":
            return execute_end_node(node, context)

        # Logic Nodes (Phase 3)
        if node.type == "logic":
            return execute_logic_node(node, context, graph, enable_debug)

        # Join Nodes (Phase 4)
        if node.type == "join":
            return execute_join_node(node, context, graph, enable_debug)

        # Analysis Nodes
        if node.type == "analysis":
            # 1. Lade Prompt (Part 3: inline_template support)
            prompt_template = await load_prompt_template(node, context)
            source_type = "inline" if node.inline_template else "reference"
            logger.debug(f"Node {node.id}: Loaded prompt from {source_type}")

            # 2. Parse question_augmentations
            questions = []
            if node.question_augmentations:
                # Convert list of dicts to JSONB-like format for parser
                questions_jsonb = [q.model_dump() if hasattr(q, 'model_dump') else q for q in node.question_augmentations]
                questions = parse_question_augmentations_from_jsonb(questions_jsonb)
                logger.debug(f"Node {node.id}: {len(questions)} question augmentations")

            # 3. Augment Prompt
            if questions:
                augmented_prompt = augment_prompt_with_questions(
                    base_prompt=prompt_template,
                    questions=questions
                )
            else:
                augmented_prompt = prompt_template

            # 4. LLM Call
            logger.debug(f"Node {node.id}: Calling LLM")
            llm_response = await openrouter_call_func(
                augmented_prompt,
                "anthropic/claude-sonnet-4"  # Default model
            )

            # 5. Parse Result Container
            parsed = parse_result_container(llm_response)
            logger.debug(f"Node {node.id}: Parsed response (status: {parsed['parsing_status']})")

            # 6. Normalize Signals
            # NOTE: decision_signals now use question.id as key (not type)
            # We need to build a catalog: id → {type, spectrum} for normalization
            normalized_signals = []
            if parsed["decision_signals"]:
                # Build catalog: id → answer_spectrum (for normalization)
                id_catalog = {}
                if questions:
                    for q in questions:
                        q_dict = q.model_dump() if hasattr(q, 'model_dump') else q
                        id_catalog[q_dict['id']] = {
                            "type": q_dict['type'],  # Keep type for normalization
                            "answer_spectrum": q_dict['answer_spectrum'],
                            "normalization_rules": None  # Node-Questions haben keine Synonyme
                        }

                # Normalize each signal (signals keyed by ID now)
                for signal_id, signal_value in parsed["decision_signals"].items():
                    if signal_id in id_catalog:
                        q_config = id_catalog[signal_id]
                        # Use the type-based catalog for normalization rules (if any)
                        type_catalog_entry = catalog.get(q_config['type'], {})

                        # Normalize with question-specific spectrum
                        normalized_signal = normalize_decision_signal(
                            question_type=signal_id,  # Use ID as question_type
                            raw_value=signal_value,
                            answer_spectrum=q_config['answer_spectrum'],
                            normalization_rules=type_catalog_entry.get('normalization_rules')
                        )

                        normalized_signals.append(normalized_signal)
                        logger.debug(f"Node {node.id}: Normalized signal '{signal_id}' = '{signal_value}' → '{normalized_signal.normalized_value}' (status: {normalized_signal.status})")

                logger.info(f"Node {node.id}: Normalized {len(normalized_signals)} signals")

            return NodeExecutionState(
                node_id=node.id,
                status=NodeStatus.EXECUTED,
                analysis_core=parsed["analysis_core"],
                decision_signals=parsed["decision_signals"],
                normalized_signals=normalized_signals,
                reasoning_anchors=parsed.get("reasoning_anchors"),
                # Debug information (nur wenn enable_debug=True)
                debug_prompt=augmented_prompt if enable_debug else None,
                debug_raw_response=llm_response if enable_debug else None,
                debug_node_type="analysis",
                debug_prompt_slug=node.prompt_slug if enable_debug else None,
                started_at=started_at,
                completed_at=datetime.utcnow().isoformat()
            )

        # Unbekannter Node-Typ (Phase 4: join)
        raise ValueError(f"Node type '{node.type}' not implemented yet (Phase 4+)")

    except Exception as e:
        logger.error(f"Node execution failed ({node.id}): {e}", exc_info=True)
        return NodeExecutionState(
            node_id=node.id,
            status=NodeStatus.FAILED,
            error=str(e),
            started_at=started_at,
            completed_at=datetime.utcnow().isoformat()
        )


def execute_logic_node(
    node,
    context: Dict[str, Any],
    graph: WorkflowGraph,
    enable_debug: bool = False
) -> NodeExecutionState:
    """
    Führt Logic Node aus (Phase 3).

    Args:
        node: WorkflowNode vom Typ "logic"
        context: Execution context mit node_results
        graph: WorkflowGraph

    Returns:
        NodeExecutionState mit evaluation_result in metadata

    Logic:
        1. Evaluiere node.condition.expression
        2. Wähle then_path oder else_path basierend auf Ergebnis
        3. Bei Fehler/Unsicherheit: Wende fallback an
        4. Markiere gewählte Edge(s) als aktiv, andere als inaktiv
        5. Return NodeExecutionState mit evaluation_result
    """
    started_at = datetime.utcnow().isoformat()

    try:
        if not node.condition:
            error_msg = f"Logic node {node.id} has no condition defined"
            logger.error(error_msg)
            return NodeExecutionState(
                node_id=node.id,
                status=NodeStatus.FAILED,
                error=error_msg,
                started_at=started_at,
                completed_at=datetime.utcnow().isoformat()
            )

        # Handle both formats (thanks to Union[LogicExpression, Condition] type):
        # 1. Direct LogicExpression (UI format): node.condition is LogicExpression
        # 2. Wrapped in Condition (legacy): node.condition is Condition with .expression
        from workflow_models import LogicExpression, Condition

        expression = None

        if isinstance(node.condition, LogicExpression):
            # UI format: direct LogicExpression
            expression = node.condition
        elif isinstance(node.condition, Condition):
            # Legacy format: wrapped in Condition
            expression = node.condition.expression
        else:
            # Fallback: try to detect format manually
            if hasattr(node.condition, 'operator') and hasattr(node.condition, 'operands'):
                expression = node.condition  # Looks like LogicExpression
            elif hasattr(node.condition, 'expression'):
                expression = node.condition.expression  # Looks like Condition

        if expression is None:
            error_msg = f"Logic node {node.id} has no valid condition/expression defined"
            logger.error(error_msg)
            return NodeExecutionState(
                node_id=node.id,
                status=NodeStatus.FAILED,
                error=error_msg,
                started_at=started_at,
                completed_at=datetime.utcnow().isoformat()
            )

        # 1. Evaluiere Bedingung
        result, error = evaluate_logic_expression(expression, context)

        if error:
            # Fehler bei Evaluation → Fallback anwenden
            logger.warning(f"Logic node {node.id}: evaluation error: {error}")
            activated_edges = _apply_fallback(node, graph, context, error)
        else:
            # Erfolgreiche Evaluation
            logger.info(f"Logic node {node.id}: evaluated to {result}")

            # 2. Wähle Pfad basierend auf Ergebnis
            if result:
                # Condition TRUE → then_path aktivieren
                activated_edges = _get_edges_by_label(node.id, "then", graph)
            else:
                # Condition FALSE → else_path aktivieren
                activated_edges = _get_edges_by_label(node.id, "else", graph)

        # 3. Markiere Edges als aktiv/inaktiv
        outgoing_edges = [e for e in graph.edges if e.from_node == node.id]
        for edge in outgoing_edges:
            context["active_edges"][edge.id] = edge.id in activated_edges

        logger.debug(f"Logic node {node.id}: activated edges: {activated_edges}")

        return NodeExecutionState(
            node_id=node.id,
            status=NodeStatus.EXECUTED,
            started_at=started_at,
            completed_at=datetime.utcnow().isoformat(),
            analysis_core=json.dumps({
                "evaluation_result": result if not error else None,
                "error": error,
                "activated_edges": activated_edges
            }),
            # Debug information
            debug_prompt=json.dumps(expression.model_dump() if hasattr(expression, 'model_dump') else str(expression), ensure_ascii=False) if enable_debug and expression else None,
            debug_raw_response=json.dumps({
                "result": result,
                "error": error,
                "activated_edges": activated_edges
            }, ensure_ascii=False) if enable_debug else None,
            debug_node_type="logic"
        )

    except Exception as e:
        logger.error(f"Logic node execution failed ({node.id}): {e}", exc_info=True)
        return NodeExecutionState(
            node_id=node.id,
            status=NodeStatus.FAILED,
            error=str(e),
            started_at=started_at,
            completed_at=datetime.utcnow().isoformat()
        )


def execute_join_node(
    node,
    context: Dict[str, Any],
    graph: WorkflowGraph,
    enable_debug: bool = False
) -> NodeExecutionState:
    """
    Führt Join Node aus (Phase 4).

    Args:
        node: WorkflowNode vom Typ "join"
        context: Execution context mit node_results
        graph: WorkflowGraph

    Returns:
        NodeExecutionState mit konsolidierten Daten

    Logic:
        1. Evaluiere Join-Strategie (via join_evaluator)
        2. Prüfe ob ready (alle erforderlichen Pfade verfügbar?)
        3. Konsolidiere Ergebnisse (merge analysis_core, combine signals)
        4. Return NodeExecutionState mit konsolidierten Daten

    Error Handling:
        - wait_all + fehlende Pfade → FAILED
        - wait_any + keine Pfade → FAILED
        - best_effort + keine Pfade → EXECUTED (mit metadata)
    """
    started_at = datetime.utcnow().isoformat()

    try:
        logger.info(f"Executing join node: {node.id}")

        # 1. Evaluiere Join-Node
        join_result = evaluate_join_node_core(node, graph, context)

        # 2. Prüfe ob ready
        if not join_result.ready:
            # Strategie nicht erfüllt → FAILED
            logger.warning(f"Join node {node.id}: Not ready - {join_result.error}")
            return NodeExecutionState(
                node_id=node.id,
                status=NodeStatus.FAILED,
                error=join_result.error,
                started_at=started_at,
                completed_at=datetime.utcnow().isoformat(),
                metadata=join_result.metadata
            )

        # 3. Baue konsolidierten analysis_core
        # Format: JSON mit node_id → analysis_core mapping
        consolidated_core_json = json.dumps(
            join_result.consolidated_analysis_core,
            ensure_ascii=False,
            indent=2
        )

        # 4. Log Konsolidierung
        executed_count = join_result.metadata.get("executed_paths", 0)
        total_count = join_result.metadata.get("total_paths", 0)
        logger.info(
            f"Join node {node.id}: Consolidated {executed_count}/{total_count} paths"
        )

        # 5. Convert consolidated_signals Dict → List[NormalizedSignal]
        # (NodeExecutionState expects List, but join_evaluator returns Dict)
        consolidated_signals_list = list(join_result.consolidated_signals.values())

        # 6. Return NodeExecutionState
        return NodeExecutionState(
            node_id=node.id,
            status=NodeStatus.EXECUTED,
            analysis_core=consolidated_core_json,
            normalized_signals=consolidated_signals_list,
            metadata=join_result.metadata,
            started_at=started_at,
            completed_at=datetime.utcnow().isoformat(),
            # Debug information
            debug_prompt=f"Join Strategy: {node.join_strategy.value if node.join_strategy else 'wait_all'}\nPaths: {total_count}\nMinimum Required: {node.min_paths if node.min_paths else 'all'}" if enable_debug else None,
            debug_raw_response=json.dumps({
                "executed_paths": executed_count,
                "total_paths": total_count,
                "strategy": node.join_strategy.value if node.join_strategy else 'wait_all',
                "consolidated_nodes": list(join_result.consolidated_analysis_core.keys())
            }, ensure_ascii=False) if enable_debug else None,
            debug_node_type="join"
        )

    except Exception as e:
        logger.error(f"Join node execution failed ({node.id}): {e}", exc_info=True)
        return NodeExecutionState(
            node_id=node.id,
            status=NodeStatus.FAILED,
            error=str(e),
            started_at=started_at,
            completed_at=datetime.utcnow().isoformat()
        )


def execute_end_node(
    node,
    context: Dict[str, Any]
) -> NodeExecutionState:
    """
    Führt End Node aus (Phase 4 - Template Engine).

    Args:
        node: WorkflowNode vom Typ "end"
        context: Execution context mit node_results

    Returns:
        NodeExecutionState mit finaler Ausgabe in analysis_core

    Output Modes:
        - AUTO: Concatenates all analysis_core values (backward compatible)
        - TEMPLATE: Renders Jinja2 template with {{node_id.property}} placeholders

    Template Context:
        - {{node_id.analysis_core}}: Analysis text from node
        - {{node_id.decision_signals}}: Dict of raw decision signals
        - {{node_id.decision_signals.key}}: Specific signal value
        - {{node_id.status}}: Node execution status
        - Conditional rendering: {% if node_id %}...{% endif %}
        - Default values: {{node_id|default("fallback")}}

    Example Template:
        ```
        # Final Analysis

        ## Body Composition
        {{body_analysis.analysis_core}}

        {% if training_analysis %}
        ## Training Recommendation
        {{training_analysis.analysis_core}}
        {% endif %}

        ## Decision Factors
        - Relevance: {{body_analysis.decision_signals.relevanz}}
        - Priority: {{body_analysis.decision_signals.prioritaet}}
        ```
    """
    started_at = datetime.utcnow().isoformat()

    try:
        logger.info(f"Executing end node: {node.id}")

        # Determine output mode (default: AUTO for backward compatibility)
        output_mode = node.output_mode or EndNodeOutputMode.AUTO

        if output_mode == EndNodeOutputMode.AUTO:
            # AUTO mode: Concatenate all analysis_core values
            logger.debug(f"End node {node.id}: Using AUTO output mode")

            # Get graph from context for node label lookup
            graph = context.get("graph")

            combined_analysis = []
            for node_id, node_state in context.get("node_results", {}).items():
                if node_state.status == NodeStatus.EXECUTED and node_state.analysis_core:
                    # Get node label from graph (fallback to node_id if not found)
                    node_label = node_id  # default
                    if graph:
                        node_obj = next((n for n in graph.nodes if n.id == node_id), None)
                        if node_obj and hasattr(node_obj, 'data') and node_obj.data:
                            node_label = node_obj.data.get('label', node_id)

                    combined_analysis.append(f"## {node_label}\n{node_state.analysis_core}")

            final_output = "\n\n".join(combined_analysis) if combined_analysis else "[No analysis generated]"

        elif output_mode == EndNodeOutputMode.TEMPLATE:
            # TEMPLATE mode: Render Jinja2 template
            logger.debug(f"End node {node.id}: Using TEMPLATE output mode")

            if not node.template:
                raise ValueError(f"End node {node.id} has output_mode=TEMPLATE but no template defined")

            # Build template context: {{node_id}} → {analysis_core, signal_X, question_X, status}
            template_context = {}
            graph = context.get("graph")  # WorkflowGraph object

            for node_id, node_state in context.get("node_results", {}).items():
                node_context = {
                    "analysis_core": node_state.analysis_core or "",
                    "decision_signals": node_state.decision_signals or {},
                    "reasoning_anchors": node_state.reasoning_anchors or "",
                    "status": node_state.status.value if node_state.status else "unknown",
                }
                # Backward-compatible shortcut:
                # allow {{node_id.relevanz}} in addition to {{node_id.decision_signals.relevanz}}
                if node_state.decision_signals:
                    for signal_key, signal_value in node_state.decision_signals.items():
                        node_context[signal_key] = signal_value

                # Add normalized signals as {{node_id.signal_ID}}
                # NOTE: question_type now IS the ID (not the type!)
                if node_state.normalized_signals:
                    for signal in node_state.normalized_signals:
                        # Convert NormalizedSignal object to dict if needed
                        signal_dict = signal.model_dump() if hasattr(signal, 'model_dump') else signal
                        q_id = signal_dict['question_type']  # This is actually the ID now!

                        signal_key = f"signal_{q_id}"
                        signal_value = signal_dict['normalized_value'] or signal_dict['raw_value']
                        node_context[signal_key] = signal_value
                        logger.info(f"Mapped signal: {q_id} → {signal_key} = '{signal_value}'")

                # Add question texts as {{node_id.question_ID}}
                if graph:
                    workflow_node = next((n for n in graph.nodes if n.id == node_id), None)
                    if workflow_node and workflow_node.question_augmentations:
                        for q in workflow_node.question_augmentations:
                            q_dict = q.model_dump() if hasattr(q, 'model_dump') else q
                            q_id = q_dict.get('id')
                            if q_id:
                                question_key = f"question_{q_id}"
                                node_context[question_key] = q_dict.get('question', '')

                template_context[node_id] = node_context

            logger.debug(f"End node {node.id}: Built template context for {len(template_context)} nodes")

            # DEBUG: Log template context keys for each node
            for node_id, node_ctx in template_context.items():
                logger.info(f"End node template context[{node_id}]: {list(node_ctx.keys())}")
                # Log signal keys specifically
                signal_keys = [k for k in node_ctx.keys() if k.startswith('signal_')]
                question_keys = [k for k in node_ctx.keys() if k.startswith('question_')]
                if signal_keys:
                    logger.info(f"  Signals: {signal_keys} → values: {[(k, node_ctx[k]) for k in signal_keys]}")
                if question_keys:
                    logger.info(f"  Questions: {question_keys} → values: {[(k, node_ctx[k]) for k in question_keys]}")

            # Render template
            try:
                # ChainableUndefined keeps missing nested attributes renderable
                # so Jinja's default filter can handle them.
                jinja_env = Environment(undefined=ChainableUndefined)
                jinja_template = jinja_env.from_string(node.template)
                final_output = jinja_template.render(template_context)
                logger.info(f"End node {node.id}: Template rendered successfully ({len(final_output)} chars)")
            except TemplateError as te:
                error_msg = f"Template rendering failed: {str(te)}"
                logger.error(f"End node {node.id}: {error_msg}")
                return NodeExecutionState(
                    node_id=node.id,
                    status=NodeStatus.FAILED,
                    error=error_msg,
                    started_at=started_at,
                    completed_at=datetime.utcnow().isoformat()
                )

        else:
            raise ValueError(f"Unknown output_mode: {output_mode}")

        # Return NodeExecutionState with final output
        return NodeExecutionState(
            node_id=node.id,
            status=NodeStatus.EXECUTED,
            analysis_core=final_output,
            started_at=started_at,
            completed_at=datetime.utcnow().isoformat()
        )

    except Exception as e:
        logger.error(f"End node execution failed ({node.id}): {e}", exc_info=True)
        return NodeExecutionState(
            node_id=node.id,
            status=NodeStatus.FAILED,
            error=str(e),
            started_at=started_at,
            completed_at=datetime.utcnow().isoformat()
        )


def _apply_fallback(
    node,
    graph: WorkflowGraph,
    context: Dict[str, Any],
    error: str
) -> List[str]:
    """
    Wendet Fallback-Strategie an.

    Args:
        node: WorkflowNode
        graph: WorkflowGraph
        context: Execution context
        error: Fehler bei Evaluation

    Returns:
        Liste von Edge-IDs die aktiviert werden sollen
    """
    if not node.fallback:
        # Default: Konservativ (else-Pfad nehmen)
        logger.warning(f"Node {node.id}: No fallback configured, using DEFAULT_PATH")
        return _get_edges_by_label(node.id, "else", graph)

    strategy = node.fallback.strategy

    if strategy == FallbackStrategy.CONSERVATIVE_SKIP:
        # Skip alle outgoing edges
        logger.info(f"Node {node.id}: CONSERVATIVE_SKIP - deactivating all paths")
        return []

    elif strategy == FallbackStrategy.DEFAULT_PATH:
        # Nimm else-Pfad
        logger.info(f"Node {node.id}: DEFAULT_PATH - taking else path")
        return _get_edges_by_label(node.id, "else", graph)

    elif strategy == FallbackStrategy.UNCERTAINTY_PATH:
        # Expliziter Unsicherheits-Pfad (falls vorhanden)
        logger.info(f"Node {node.id}: UNCERTAINTY_PATH")
        uncertainty_edges = _get_edges_by_label(node.id, "uncertainty", graph)
        if uncertainty_edges:
            return uncertainty_edges
        else:
            # Fallback to else
            logger.warning(f"Node {node.id}: No uncertainty path found, using else")
            return _get_edges_by_label(node.id, "else", graph)

    elif strategy == FallbackStrategy.DOCUMENT_ONLY:
        # Alle Pfade aktiv lassen (wie ohne Bedingung)
        logger.info(f"Node {node.id}: DOCUMENT_ONLY - all paths remain active")
        outgoing_edges = [e for e in graph.edges if e.from_node == node.id]
        return [e.id for e in outgoing_edges]

    else:
        logger.warning(f"Node {node.id}: Unknown fallback strategy {strategy}, using DEFAULT_PATH")
        return _get_edges_by_label(node.id, "else", graph)


def _get_edges_by_label(node_id: str, label: str, graph: WorkflowGraph) -> List[str]:
    """
    Findet alle ausgehenden Edges mit bestimmtem Label.

    Unterstützt beide Formate:
    - Legacy: e.label == label (z.B. "then", "else")
    - UI: e.source_handle == label (z.B. "true", "false")

    Args:
        node_id: Node-ID
        label: Edge-Label oder sourceHandle (z.B. "then"/"true", "else"/"false")
        graph: WorkflowGraph

    Returns:
        Liste von Edge-IDs
    """
    # Map label to sourceHandle equivalents
    label_to_handle = {
        "then": "true",
        "else": "false"
    }
    handle_equivalent = label_to_handle.get(label, label)

    matching_edges = [
        e.id for e in graph.edges
        if e.from_node == node_id and (
            e.label == label or  # Legacy format
            (hasattr(e, 'source_handle') and e.source_handle == handle_equivalent)  # UI format
        )
    ]
    return matching_edges


def _has_active_incoming_edge(node, graph: WorkflowGraph, context: Dict[str, Any]) -> bool:
    """
    Prüft ob Node mindestens eine aktive incoming edge hat.

    Args:
        node: WorkflowNode
        graph: WorkflowGraph
        context: Execution context mit active_edges

    Returns:
        True wenn mindestens eine incoming edge aktiv ist
    """
    # Start-Node hat keine incoming edges → immer aktiv
    if node.type == "start":
        return True

    incoming_edges = [e for e in graph.edges if e.to_node == node.id]

    # Keine incoming edges → nicht erreichbar
    if not incoming_edges:
        return False

    # Prüfe ob mindestens eine aktiv ist
    active_edges = context.get("active_edges", {})
    for edge in incoming_edges:
        if active_edges.get(edge.id, True):  # Default: True (Phase 2 Kompatibilität)
            return True

    return False


async def load_prompt_template(node: WorkflowNode, context: Dict[str, Any]) -> str:
    """
    Lädt Prompt-Template aus DB (reference mode) oder direkt vom Node (inline mode).

    Part 3: Inline Prompts - Unterstützt zwei Modi:
    - Reference Mode: prompt_slug → Template aus ai_prompts Tabelle
    - Inline Mode: inline_template → Template direkt vom Node

    Args:
        node: WorkflowNode mit prompt_slug ODER inline_template
        context: {"variables": {"name": "Lars", ...}, "profile_id": "..."}

    Returns:
        Resolved prompt template

    Raises:
        HTTPException: Wenn weder prompt_slug noch inline_template gesetzt

    Beispiel:
        >>> node = WorkflowNode(id="n1", prompt_slug="pipeline_body")
        >>> template = await load_prompt_template(node, {"profile_id": "123"})
        >>> "{{name}}" not in template
        True
    """
    from placeholder_resolver import get_placeholder_example_values, get_placeholder_catalog
    from prompt_executor import resolve_placeholders
    from fastapi import HTTPException

    # Mode 1: Inline Template (NEU)
    if node.inline_template:
        logger.debug(f"Node {node.id}: Using inline template ({len(node.inline_template)} chars)")
        template = node.inline_template

    # Mode 2: Reference (bestehend)
    elif node.prompt_slug:
        logger.debug(f"Node {node.id}: Loading prompt '{node.prompt_slug}' from DB")
        with get_db() as conn:
            cur = get_cursor(conn)
            cur.execute(
                "SELECT template FROM ai_prompts WHERE slug = %s AND active = true",
                (node.prompt_slug,)
            )
            row = cur.fetchone()
            if not row:
                raise HTTPException(status_code=404, detail=f"Prompt not found: {node.prompt_slug}")
            template = row['template']

    # Mode 3: Error - weder inline noch reference
    else:
        raise HTTPException(
            status_code=400,
            detail=f"Node {node.id}: Either prompt_slug or inline_template required"
        )

    # Resolve Placeholders using modern prompt_executor method
    profile_id = context.get("profile_id")
    if not profile_id:
        raise HTTPException(status_code=400, detail="profile_id required in context")

    # Build variables dict with ALL registered placeholders
    variables = {}

    try:
        # Get all placeholder values from registry
        processed_placeholders = get_placeholder_example_values(profile_id)
        logger.info(f"🔍 DEBUG: Loaded {len(processed_placeholders)} placeholders from registry")
        logger.info(f"🔍 DEBUG: Sample keys (first 3): {list(processed_placeholders.keys())[:3]}")

        # Remove {{ }} from keys (placeholder_resolver returns them with wrappers)
        cleaned_placeholders = {
            key.replace('{{', '').replace('}}', '').strip(): value
            for key, value in processed_placeholders.items()
        }
        logger.info(f"🔍 DEBUG: Cleaned keys (first 3): {list(cleaned_placeholders.keys())[:3]}")
        logger.info(f"🔍 DEBUG: Sample values: name={cleaned_placeholders.get('name')}, age={cleaned_placeholders.get('age')}, geschlecht={cleaned_placeholders.get('geschlecht')}")

        variables.update(cleaned_placeholders)
    except Exception as e:
        logger.error(f"❌ CRITICAL: Failed to load placeholders for workflow: {e}", exc_info=True)

    # Add workflow node outputs as placeholders (Part 3: Inline Prompts)
    # Format: node_id.analysis_core, node_id.signal_xyz, node_id.question_xyz
    node_results = context.get("node_results", {})
    if node_results:
        logger.info(f"🔍 DEBUG: Adding {len(node_results)} node outputs as placeholders")
        for node_id, node_state in node_results.items():
            # analysis_core
            if hasattr(node_state, 'analysis_core') and node_state.analysis_core:
                key = f"{node_id}.analysis_core"
                variables[key] = node_state.analysis_core
                logger.debug(f"  Added placeholder: {key} = {node_state.analysis_core[:50]}...")

            # decision_signals (keyed by question ID)
            if hasattr(node_state, 'decision_signals') and node_state.decision_signals:
                for signal_id, signal_value in node_state.decision_signals.items():
                    # Signal placeholder: node_id.signal_question_id
                    signal_key = f"{node_id}.signal_{signal_id}"
                    variables[signal_key] = signal_value
                    logger.debug(f"  Added placeholder: {signal_key} = {signal_value}")

            # Question texts (from graph metadata if available)
            # NOTE: Question text placeholders are populated from graph in PlaceholderPicker
            # Here we only add if available in node_state metadata
            if hasattr(node_state, 'metadata') and isinstance(node_state.metadata, dict):
                questions = node_state.metadata.get('questions', [])
                for q in questions:
                    if isinstance(q, dict) and 'id' in q and 'question' in q:
                        question_key = f"{node_id}.question_{q['id']}"
                        variables[question_key] = q['question']
                        logger.debug(f"  Added placeholder: {question_key}")

    # Load catalog for |d modifier support
    try:
        catalog = get_placeholder_catalog(profile_id)
    except Exception as e:
        catalog = None
        logger.warning(f"Failed to load placeholder catalog for workflow: {e}")

    logger.info(f"🔍 DEBUG: Template before resolution:\n{template[:200]}...")
    logger.info(f"🔍 DEBUG: Variables dict has {len(variables)} entries")

    # Resolve with modern executor
    debug_info = {}
    resolved = resolve_placeholders(
        template=template,
        variables=variables,
        debug_info=debug_info,
        catalog=catalog
    )

    logger.info(f"🔍 DEBUG: Resolved placeholders: {debug_info.get('resolved_placeholders', {})}")
    logger.info(f"🔍 DEBUG: Unresolved placeholders: {debug_info.get('unresolved_placeholders', [])}")
    logger.info(f"🔍 DEBUG: Template after resolution:\n{resolved[:200]}...")

    return resolved




def aggregate_results(node_states: List[NodeExecutionState], graph) -> Dict[str, Any]:
    """
    Aggregiert Ergebnisse aller Knoten.

    Phase 4: Wenn End Node existiert, verwende NUR dessen Output.
    Sonst (backward compatible): Kombiniere alle analysis_core Werte.

    Args:
        node_states: Liste aller NodeExecutionState
        graph: WorkflowGraph object (to identify End Nodes)

    Returns:
        {
            "analysis_core": "Final output (from End Node or combined)",
            "all_signals": [{question_type, normalized_value, status}, ...],
            "total_nodes": 3,
            "executed_nodes": 3,
            "failed_nodes": 0
        }

    Beispiel:
        >>> states = [
        ...     NodeExecutionState(node_id="n1", status=NodeStatus.EXECUTED, analysis_core="Test 1"),
        ...     NodeExecutionState(node_id="n2", status=NodeStatus.EXECUTED, analysis_core="Test 2"),
        ...     NodeExecutionState(node_id="end", status=NodeStatus.EXECUTED, analysis_core="Final")
        ... ]
        >>> result = aggregate_results(states, graph)
        >>> result["analysis_core"]
        "Final"
    """
    # Find End Node in graph
    end_node_ids = [n.id for n in graph.nodes if n.type == "end"]

    combined_analysis = []
    all_signals = []
    final_output = None

    for state in node_states:
        # Check if this is an End Node
        is_end_node = state.node_id in end_node_ids

        if state.status == NodeStatus.EXECUTED and state.analysis_core:
            if is_end_node:
                # End Node output is the final output (don't add to combined_analysis)
                final_output = state.analysis_core
            else:
                # Regular node - add to combined analysis
                # Get node label from graph (fallback to node_id)
                node_label = state.node_id  # default
                node_obj = next((n for n in graph.nodes if n.id == state.node_id), None)
                if node_obj and hasattr(node_obj, 'data') and node_obj.data:
                    node_label = node_obj.data.get('label', state.node_id)

                combined_analysis.append(f"## {node_label}\n{state.analysis_core}")

        if state.normalized_signals:
            all_signals.extend([s.model_dump() for s in state.normalized_signals])

    # Phase 4: If End Node exists, use its output ONLY
    # Backward compatible: If no End Node, use combined_analysis
    if final_output:
        primary_output = final_output
    else:
        primary_output = "\n\n".join(combined_analysis)

    return {
        "analysis_core": primary_output,
        "combined_analysis": "\n\n".join(combined_analysis),  # Keep for debugging (excludes End Node)
        "all_signals": all_signals,
        "total_nodes": len(node_states),
        "executed_nodes": sum(1 for s in node_states if s.status == NodeStatus.EXECUTED),
        "failed_nodes": sum(1 for s in node_states if s.status == NodeStatus.FAILED),
        "skipped_nodes": sum(1 for s in node_states if s.status == NodeStatus.SKIPPED)  # Phase 3
    }


def save_execution_state(
    execution_id: str,
    workflow_id: Optional[str],  # None when using graph_data directly (Phase 5)
    profile_id: str,
    node_states: List[NodeExecutionState],
    status: str,
    started_at: str,
    completed_at: Optional[str] = None,
    error: Optional[str] = None
):
    """
    Speichert Execution State in workflow_executions.

    Args:
        execution_id: UUID der Execution
        workflow_id: UUID des Workflows (None wenn graph_data direkt verwendet wird)
        profile_id: UUID des Profils
        node_states: Liste aller NodeExecutionState
        status: 'completed' | 'failed' | 'partial'
        started_at: ISO timestamp
        completed_at: ISO timestamp (optional)
        error: Fehlermeldung (optional)

    Beispiel:
        >>> save_execution_state(
        ...     execution_id="exec-123",
        ...     workflow_id="wf-456",
        ...     profile_id="prof-789",
        ...     node_states=[],
        ...     status="completed",
        ...     started_at="2026-04-03T12:00:00"
        ... )
    """
    # Serialize node_states to JSON
    node_states_json = [s.model_dump() for s in node_states]

    with get_db() as conn:
        cur = get_cursor(conn)
        cur.execute("""
            INSERT INTO workflow_executions
            (id, workflow_id, profile_id, status, node_states, execution_log, started_at, completed_at)
            VALUES (%s, %s, %s, %s, %s, %s, %s, %s)
            ON CONFLICT (id) DO UPDATE SET
                status = EXCLUDED.status,
                node_states = EXCLUDED.node_states,
                execution_log = EXCLUDED.execution_log,
                completed_at = EXCLUDED.completed_at
        """, (
            execution_id,
            workflow_id,
            profile_id,
            status,
            json.dumps(node_states_json),
            json.dumps({"error": error} if error else {}),
            started_at,
            completed_at
        ))
        conn.commit()

    logger.info(f"Saved execution state: {execution_id} (status: {status})")